“I don't believe Claude Code will exist in its current form in six months” Benۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗ☁️, CEO & Co-Founder Freestyle, thinks coding agents are moving from local machines to the cloud, where a single task could get the attention of 20+ agents at once. Each gets a complete copy of... your stack, production environment included. Each can spend a week testing and refining its approach. You compare the results and take the best one forward. The economics aren’t there yet. Ben expects cost per task to fall another 99% over the next four years, making that level of parallel work practical. Freestyle builds the computers for it: full Linux VMs for tasks that run for hours, days, or weeks. Clone a running machine, memory included, and let agents pursue different approaches from the same starting point. Pause and resume with their state intact. Inside Freestyle, Ben already gives agents a week to find improvements to its VM technology. Roughly 700 tests and 90 metrics help the team judge whether the work made things better. He calls this goal engineering: define the outcome clearly, give agents time to work toward it, and measure whether they’re making progress. "Full Linux VMs for AI Agents" Built for complex tasks that run for hours, days, or weeks 🎙️ Benۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗۗ☁️, CEO & Co-Founder, Freestyle on Fondo.com START 00:57 Why coding agents could move from local to the cloud 02:22 From harness engineering to goal engineering 03:08 Giving agents a week to improve measurable results 04:13 Why defining the problem becomes the bottleneck 04:48 How early access to o1 changed Freestyle’s direction 05:51 Why frustrated sandbox users revealed a bigger opportunity 07:28 Giving agents a computer instead of building custom tools 08:12 Getting into YC 11:17 Snapshotting VMs and running agents in parallel 12:24 The economics of letting multiple agents attempt every task 13:26 How GPU supply could drive down agent costs 15:09 Try Freestyle and follow Benshow more

@benswerd @freestyle_dev local agent on one machine was always a temporary shape. once the task can fork twenty sandboxes the product has to be the control plane.

@benswerd @freestyle_dev local agents have one fatal flaw and it is called closing your laptop

@benswerd @freestyle_dev the freestyle detail that stuck with me: 700 tests and 90 metrics before they call a branch the winner. without a scoreboard like that, twenty cloud agents just spend money in parallel.

@benswerd @freestyle_dev if those twenty can run the 700 tests, a week mostly rewards whoever games the suite fastest. hold out a slice none of them can see, score on that, and the bake-off measures fixes instead of tests.

@benswerd @freestyle_dev local agent to cloud swarm in six months. that timeline is wild

@benswerd @freestyle_dev cloud agents make sense to me. mine kept dying when my laptop slept mid run, so i moved them off my machine and stopped babysitting them.

@benswerd @freestyle_dev This turns coding into a bake-off: give isolated agents the same starting machine and measurable goal, then promote the approach that actually improves the tests.

@benswerd @freestyle_dev Great interview. Constantly putting insights from papers into production sounds fun!

@benswerd @freestyle_dev 20 agents on one task means 19 losing branches you still pay for. the 700 tests and 90 metrics are the real story here. without a scoreboard like that, parallel agents just multiply the bill. the number that has to fall 99% is cost per accepted result, not cost per run.

@benswerd @freestyle_dev 20 agents on one task makes the bill question sharper, not smaller. if every task fans out to 20 copies, someone has to see cost per merged change, or parallel just means paying 20x for the one that won

@benswerd @freestyle_dev Twenty agents can explore more paths, but twenty copies of the same context can also repeat the same blind spot. The useful control is independent verification and one clear stop condition before any agent acts outside the sandbox.

@benswerd @freestyle_dev Giving 20 agents a full week to disagree sounds like either the future of software or the world’s most expensive group chat. The VM + 700 tests detail makes the idea feel a lot less sci-fi.

@benswerd @freestyle_dev six months is a bold call. did he say what he thinks replaces it?

@benswerd @freestyle_dev spinning up 20 copies of production is the part nobody has solved. containers are fine, it's the data that won't clone, and a stack you can't rebuild from scratch caps the whole thing.